Improving Cross-Domain Concept Detection via Object-Based Features

Learned visual concept models often do not work well for other domains not considered during training, because a concept's visual appearance strongly depends on the domain of the corresponding image or video source. In this paper, a novel approach to improve cross-domain concept detection is presented. The proposed approach uses features based on object detection results in addition to Bag-of-Visual-Words features as inputs to concept classifiers. Experiments conducted on TRECVid videos using a high-performance computing cluster show that the additional use of object-based features significantly improves the generalization properties of the learned concept models in cross-domain settings, for example, from broadcast news videos to documentary films and vice versa.